Google researchers launched a new framework for generating long videos

Google researchers introduced a set of multi-agent frameworks designed to automate the creation of long-form, coherent videos. By using an orchestration layer built on Gemini and Veo, the system manages narrative structure and visual consistency, which often fail in standard video models due to character or background drift. The researchers demonstrated the capability by generating a continuous ten-minute video sequence.

The framework operates through a hierarchy of agents that handle different tasks. An orchestrator uses a multi-armed bandit algorithm to select creative strategies, while specialized agents manage storyboarding, motion, and audio. Another component, CANVAS, maintains visual memory to ensure characters and objects remain consistent across multiple shots, while VQQA provides automated feedback to fix compositional artifacts.

The system is not entirely automated and requires human creative input to set the initial trajectory. While it successfully mitigates common errors like background shifts, it operates as a prompt optimizer and does not provide pixel-level control. The research does not currently show how the system handles complex interactive scenarios or extreme artistic variations.